What are differences between Hailo-8 and Hailo-8L?


What are differences between Hailo-8 and Hailo-8L?
Actually, I can find TOPS information only. (26 vs. 13)
Are there any details about chipset specifications such as tensor core counts, memory size and core clocks?

And they seem to use different compiled hef binaries for same models, hef binaries are not compatible each other even though the model has same parameter size like the following.

|hw_arch|Network Name|Accuracy (top1)|Quantized|FPS (Batch Size=1)|FPS (Batch Size=8)|Input Resolution (HxWxC)|Params (M)|OPS (G)|

So, do we need to do separate whole “model build process” for each arch if we build models or only need to do “Allocate Resources and Generate Compiled Binary File” step in “model build process” flow?

And Is there easy way to convert hef files from Hailo-8 to Hailo-8L?


Welcome to the Hailo Community!

The Hailo-8L has half the resources of the Hailo-8. Otherwise it runs at the same speed. The Hailo-8L modules are available as M.2 key B+M and A+E which both have 2 PCIe lanes while the Hailo-8 additionally is available as M.2 key M with 4 PCIe lanes.

In your conversion script you only need to change the architecture flag to hailo8l.

This is the difference - Hailo-8L has half the amount of compute elements, the rest is identical.

You cannot convert directly a HEF from Hailo-8 to Hailo-8L, or vice-versa. You will need to compile it in advance to target device.

The conversion flow is basically built from 3 steps:

  1. Parsing
  2. Optimization (Quantization)
  3. Compilation

From that point of view, only the last step need to be re-done between Hailo-8 and Hailo-8L.

Thanks for your explanations.

You mean only the last step, “Allocate resources and Generate Compiled Binary File” in below picture is required ?

But I can see it is not recommended to change the hw_arch after Optimization in “4.4.1. Hailo Compilation Example from Hailo Archive Quantized Model to HEF” section of “Hailo Dataflow Compiler User Guide” as you see below picture.

Is it okay to change the hw_arch to Hailo-8L if the quantized model har file comes from Hailo-8?

Yes, if you want the most robust flow, is to start afresh. For the case of Hailo-8 vs. Hailo-8L, it will be ok.